The difference observed on the remote testbed doesn't show up in all-localhost testing:
HEAD [SCAN], Operations, 949967 [SCAN], AverageLatency(us), 23847.456127423375 [SCAN], MinLatency(us), 625 [SCAN], MaxLatency(us), 1806981 [SCAN], 95thPercentileLatency(ms), 56 [SCAN], 99thPercentileLatency(ms), 71 HEAD~50: 5f853cb... HBASE-11436 [SCAN], Operations, 949937 [SCAN], AverageLatency(us), 23844.437741660764 [SCAN], MinLatency(us), 961 [SCAN], MaxLatency(us), 1843125 [SCAN], 95thPercentileLatency(ms), 55 [SCAN], 99thPercentileLatency(ms), 70 0ca0ced Update CHANGES.txt for 0.98.3RC1 [SCAN], Operations, 950224 [SCAN], AverageLatency(us), 24303.889086152318 [SCAN], MinLatency(us), 956 [SCAN], MaxLatency(us), 2091141 [SCAN], 95thPercentileLatency(ms), 56 [SCAN], 99thPercentileLatency(ms), 71 I have the testbed for one more day. I'll try an educated guess. Otherwise, will need to change my vote back to +1 because I can't veto a RC for something I can't verify. On Thu, Jul 17, 2014 at 5:51 PM, Andrew Purtell <apurt...@apache.org> wrote: > Bisecting now. > > I plan to find and revert the culprit and any related commits, confirm > improvement with workload E, push those changes back onto the pile for .5, > and roll .4 RC1 on or before Monday. Phoenix has a release deadline at the > end of the month and changes in .4 they need (if not the issue). > > > On Thu, Jul 17, 2014 at 5:12 PM, Andrew Purtell <apurt...@apache.org> > wrote: > >> See thread on dev@ titled "Comparing the performance of 0.98.4 RC0 and >> 0.98.0 using YCSB - 23% perf regression in workload E" >> >> -1 on this RC for now, pending reproduction and further analysis on a dev >> box. >> >> [...] >> >> These tests were run with no security coprocessors installed, using HFile >> V2. The workload E results are a concern. *It appears we have a 23% >> decline in measured scan throughput and an 23% increase in average op time >> from 27 ms to 35 ms. *This does not correspond to any active security >> feature (though that could worsen results potentially, untested) so is >> something changed in core code. Other workloads are not affected so this is >> something specific to scanning. Perhaps delete tracking. >> >> [...] >> >> *Workload E* >> >> >> >> >> >> >> [OVERALL] RunTime(ms)16009102078826 [OVERALL]Throughput(ops/sec) 63084835 >> [INSERT] Operations499131500322 [INSERT]AverageLatency(us) 1417[INSERT] >> MinLatency(us)55 [INSERT]MaxLatency(us) 506079564468[INSERT] >> 95thPercentileLatency(ms)0 0[INSERT]99thPercentileLatency(ms) 00 [SCAN] >> Operations9500869 9499678[SCAN] AverageLatency(us) >> >> >> 2663634620 [SCAN]MinLatency(us) 746755[SCAN] MaxLatency(us)8067864 >> 4615914[SCAN]95thPercentileLatency(ms) 117136 [SCAN] >> 99thPercentileLatency(ms)169 187 >> >> >> >> >> >> On Mon, Jul 14, 2014 at 8:38 PM, Andrew Purtell <apurt...@apache.org> >> wrote: >> >>> The 1st HBase 0.98.4 release candidate (RC0) is available for download >>> at http://people.apache.org/~apurtell/0.98.4RC0/ and Maven artifacts >>> are also available in the temporary repository >>> https://repository.apache.org/content/repositories/orgapachehbase-1026/ >>> >>> Signed with my code signing key D5365CCD. >>> >>> The issues resolved in this release can be found here: >>> https://issues.apache.org/jira/secure/ReleaseNote.jspa?projectId=12310753&version=12326810 >>> >>> >>> Please try out the candidate and vote +1/-1 by midnight Pacific Time >>> (00:00 -0800 GMT) on July 21 on whether or not we should release this as >>> 0.98.4. Three +1 votes from PMC will be required to release. >>> >>> > > > -- > Best regards, > > - Andy > > Problems worthy of attack prove their worth by hitting back. - Piet Hein > (via Tom White) > -- Best regards, - Andy Problems worthy of attack prove their worth by hitting back. - Piet Hein (via Tom White)